16,483 karma · joined March 9, 2012
2010-2019 I lived in China.
Before that I lived in London, where I was born.
My personal web site: https://encona.com/
Random tools: https://tools.encona.com/
Other random stuff about me:
- English is my native language, but I also speak (to varying degrees of fluency) Mandarin, Gujarati and Spanish.
- My career so far has been a mix of startups/cofounding, big tech (I was a PM at Amazon and at Google), and exec roles.
- I'm a qualified accountant (CIMA), have two undergrad degrees (one in CS) and an MBA (sorry!).
- I got married in 2012 and our son was born in 2016 (I spend a lot of time thinking about how best to raise him to be a capable and kind adult)
Email: rahim AT encona DOT com
LinkedIn: http://www.linkedin.com/in/rahimnathwani
X: https://x.com/rahimNathwani/
To me it seems like:
- Jev should be good at telling us 'what is'
- Jev isn't trying to tell us 'what should be'
e.g. if you want your software to assume each human life has the same value, then you should rely on Jev to tell you how many human lives would be lost in each scenario, and then your software can pick. But Jev is not trained on your value system, which is neither objective truth nor something that you can derive from the inputs.
Thank you!
uv tool install mlx-audio
mlx_audio.stt.generate \
--model ibm-granite/granite-speech-5.0-470m-turboctc \
--audio.wav \
--output-path transcript
(Until now I've been using mlx-community/parakeet-tdt-0.6b-v3)It's a shame that iPhoneOS, iPadOS and Omarchy are designed to be single-user devices.
Android phones (at least Pixels) support multiple users, each with their own apps and data.
From page 4:
Problem difficulty increased across the experiment from one-step1 (problems 1–4) to two-step2 (problems 5–8) to three-step3 (problems 9–12).
From page 6: Problems increased in difficulty from one-step (problems 4-6) to two-step (problems 7-10) to three-step (problems 11-14) problems.
The fact that the AI group did worse on the harder problems than the non-AI group did on the easier ones doesn't tell us anything.- BM25 only
- Jev only
- LLM + Jev
The latter worked best.
But I wonder whether LLM on its own might have gotten close. If you're going to use an LLM for each call then it seems reasonable to test whether or not adding Jev improves performance.
EDIT: I should have had more coffee before writing this comment. As make3 points out below, I forgot that in the LLM+Jev setup, the LLM only needs to run once per query, whereas in the LLM-only setup the LLM would run once per item/segment being evaluated.
The evidence for reduced persistence is kinda weak.
The strongest experimental result in the paper is: if people use AI to help solve a series of problems, and then the AI is suddenly taken away, they perform worse on subsequent problems than people who hadn't been using AI for the earlier questions.
But:
- They're looking at really really short-term effects, like seconds or minutes. That's not really interesting if we care about long term effects (which is what the abstract says is important).
- What they found is totally expected and isn't something I'd worry about: it's kind of expected that if one guy warms up with 10 questions done himself, and the other guy warms up with 10 questions delegated to AI, then the second guy is going to do worse on the final 10 questions (because he didn't warm up).
But if you ignore the three slides about medicine, the rest of it is interesting, reasonable and informative.
An owner can derive value from the house in one of three ways:
A) Occupy the whole place.
B) Receive rent from a tenant.
C) Something in between (e.g. airbnb one room).
When property prices double, it's likely that rents roughly double. This may not increase the value the owner can derive from (A) but it will sure change the value they can derive from (B).
So they might decide to switch from living in the place to renting it out. If they don't decide to switch, then it indicates the value they derive from living in the property has either gone up, or it was always way higher than the market rent.
It seems like market rent is a reasonable benchmark to use to calculate property taxes, and property value is a reasonable proxy for market rent.
Otherwise it would be really confusing when you switch from your external keyboard to the laptop keyboard.
Spend 5 mins watching the first 5 videos on the announcement thread or github README:
https://x.com/lucasmeijer/status/2108248641605706006
https://github.com/lucasmeijer/AgentsInTheCloud/
The things that make me excited to try it:
- really nice isolation between workspaces (a docker container per workspace, and defaulting to no network comms between workspaces)
- each workspace can use whichever coding agent you want
- nice web UX which improves coding agents' own UI (not only the markdown view of agent output, but also the side panel with web browser, VS code, terminal)
We live in San Francisco so there are only a few days a year when we need a fan. Like this weekend.
From my order history:
Prominence Home Saybrook, 52 inch Indoor Farmhouse LED Ceiling Fan with Light, Pull Chain, Three Mounting Options, Dual Finish Blades, Reversible Motor - 51592-01 (Matte Nickel)
Sold by: Amazon.com
Supplied by: Other
$99.22I bought some high CRI LED bulbs from some place recommended in an HN post.
A) Structured outputs from LLMs (doesn't need fine tuning but can be expensive)
B) Classification output from fine-tuned BERT-like or GLiNER models (is calibrated well and has cheap/fast inference)
What Jev did is combine the advantages of both A and B into one model/product, and create a really good API.
They claim that a key innovation is how they've trained the model using what they call RLCD (RL from calibrated decisions). So it's not just that you can get the outputs (which is easy to add to any LLM) but that the different primitives they expose (Choice, Score, Noul) have each been calibrated. For example, they claim that if you use the Score primitive (which gives you probabilities along a bunch of choices representing a continuum) that's not just using the more general 'Choice' primitive under the hood. It's been calibrated separately.
I don't know how many of the Jev-like things we've seen do that. For example Cloudflare offers a Jev-like model with the same API, and which they say was trained with RLCD. But I don't know whether Choice and Score are different under the hood, or whether Score is just sugar on top of Choice. (Should be easy to test this, but I haven't done it.)
If someone in France is not free to hold a sign that says "Get lost, you sad jerk" then France does not have free speech. That sign contains only words. It does not create any threat. It does not incite anyone to take illegal and/or violent actions.
Regarding the Bardot case you say "I don't think this kind of exception is fairly applied to discredit what I wrote" but:
- that's not the only case I cited, and
- even in the Bardot case, what's the rationale for suppressing speech expressing an opinion? it seems she was prosecuted because her expression of an opinion might cause others to adopt that same opinion. She wasn't inciting violence or illegal acts.
- Hervé Eon Sarkozy fine
- Éric Zemmour 3000 fine
- bardot degenerates 20000
- Je me sens Charlie Coulibaly fine
None of these were threats or incitement to violence. Just insults.
And you talk about a constitutional right to free speech since the 18th century, but mocking Catholicism could be prosecuted until 1881 (i.e. in the 19th century).
3.5 Flash-Lite
3.8 Flash
3.1 Pro
In both a paid Google Workspace account (without the AI addon) and a free 'GSuite' account, I see: 3.6 Flash
3.6 Thinking
3.1 ProIn case anyone wants to make it, you can google 'venus kusudama'
But if you must know what they do other than analysis and writing, here's a recent example: https://www.educationprogress.org/p/were-suing-the-makers-of...
Disclaimer: I think Jack, Thomas and the others are doing useful work. I have contributed two articles to their Substack. I don't think the number of people involved in an organization should determine how seriously we take that organization.